Running Couplings and Dimensional Transmutation

A running coupling is a coordinate along a fixed-bare renormalization-group trajectory. For one coupling, the trajectory can be integrated explicitly: a negative cubic beta function weakens the coupling toward the ultraviolet, while a positive cubic beta function drives it toward a formal ultraviolet Landau scale. In a massless asymptotically free theory, the integration constant can be written as a mass scale $\Lambda$. Specifying $g(\mu_0)$ at a reference scale and specifying $\Lambda$ are then equivalent boundary conditions.

This replacement is dimensional transmutation. It does not make a strong-coupling mass perturbatively calculable: perturbation theory controls the relation between $g(\mu)$ and $\Lambda$ only while $g(\mu)$ is small. Scheme, coupling normalization, beta-function order, thresholds, and the stopping rule must travel with every quoted value of $\Lambda$.

Required background. Beta Functions, Running Masses, and Field Anomalous Dimensions defines $\beta(g)=\mu\,dg/d\mu\rvert_0$ and explains how its coefficients are extracted.

Helpful background. Asymptotic Scales, Remainders, Uniformity, and Optimal Truncation clarifies why an exact solution of a truncated beta function is still only a perturbative approximation.

One cubic term, two flow directions

Set

$$t\equiv\ln\frac{\mu}{\mu_0}, \qquad g(0)=g_0>0,$$

and consider

$$\beta(g)=\sigma b_0g^3, \qquad b_0>0, \qquad \sigma=\pm1.$$

Because

$$\frac{d}{dt}\frac{1}{g^2} = -\frac{2\beta(g)}{g^3} = -2\sigma b_0,$$

the solution is

$$\boxed{ \frac{1}{g^2(\mu)} = \frac{1}{g_0^2} -2\sigma b_0\ln\frac{\mu}{\mu_0} }$$

or, on the branch connected continuously to $g_0$,

$$g^2(\mu) = \frac{g_0^2}{ 1-2\sigma b_0g_0^2\ln(\mu/\mu_0) }.$$

The two signs describe different perturbative directions:

Sign and weak-coupling behavior	Formal endpoint of the one-loop solution	Sound conclusion
$\sigma=-1$: $g(\mu)\to0$ as $\mu\to\infty$	$\Lambda_{\mathrm{AF}}=\mu_0\exp[-1/(2b_0g_0^2)]$ toward the infrared	The ultraviolet is asymptotically free; stop the perturbative evolution when $g$ becomes too large
$\sigma=+1$: $g(\mu)\to0$ as $\mu\to0$	$\Lambda_{\mathrm{L}}=\mu_0\exp[+1/(2b_0g_0^2)]$ toward the ultraviolet	The ultraviolet trajectory leaves weak coupling; the truncated pole does not determine the exact ultraviolet completion

Calling either endpoint a “pole” describes the rational one-loop solution. It does not prove that a physical observable has a singularity there. For $\sigma=-1$, the infrared endpoint instead identifies the exponentially small scale at which the weak-coupling expansion loses control. For $\sigma=+1$, it gives a formal ultraviolet stopping estimate.

The sign test is local but useful. If $\beta(g)<0$ for small positive $g$, increasing $\mu$ decreases the coupling. Reversing the definition of $t$ without also reversing the flow equation swaps ultraviolet and infrared and produces a false interpretation even when the algebraic solution looks plausible.

The two-term asymptotically free flow

Now retain the first two nonzero terms,

$$\beta(g) = -b_0g^3-b_1g^5 +O(g^7), \qquad b_0>0.$$

Let

$$x\equiv g^2, \qquad c\equiv\frac{b_1}{b_0}.$$

If the displayed two-term beta function is treated as a differential equation, then

$$\frac{dx}{dt} = -2b_0x^2(1+cx).$$

Define

$$F(x) = \frac{1}{x} +c\ln\left(\frac{x}{1+cx}\right).$$

Direct differentiation gives $F'(x)=-1/[x^2(1+cx)]$, so the implicit solution is

$$\boxed{ F\bigl(g^2(\mu)\bigr) -F(g_0^2) = 2b_0\ln\frac{\mu}{\mu_0}. }$$

For $b_1=0$ this reduces to the cubic solution. If $b_1<0$, the two-term polynomial also has a positive zero

$$g_*^2=-\frac{b_0}{b_1}.$$

A trajectory with $0<g<g_*$ approaches this zero toward the infrared. That is evidence for a perturbative infrared fixed point only when $g_*$ is parametrically small and omitted terms remain controlled. A zero at order-one coupling is not upgraded to an exact fixed point by integrating the truncated equation exactly.

The two-term solution and its asymptotic expansion are derived in Collins 1984/2023, § 7.5.1, pp. 188–189. The important ordering is: truncate the beta function to a declared loop order, solve or expand it consistently, and estimate the effect of the first omitted coefficient. “Exact two-loop running” means exact integration of that truncated ordinary differential equation, not an exact statement about the QFT.

Constructing the RG-invariant scale

For any one-coupling branch on which $\beta(g)\ne0$, choose a conventional reference value $g_c$ and define

$$\Lambda_c = \mu \exp\left[ -\int_{g_c}^{g(\mu)}\frac{dg'}{\beta(g')} \right].$$

Its logarithmic derivative vanishes:

$$\frac{d\ln\Lambda_c}{d\ln\mu} = 1- \frac{1}{\beta(g)} \frac{dg}{d\ln\mu} =0.$$

Changing $g_c$ multiplies $\Lambda_c$ by a constant. Thus RG invariance does not by itself choose the normalization of the scale.

For the one-term asymptotically free beta function,

$$\boxed{ \Lambda_{(1)} = \mu\exp\left[-\frac{1}{2b_0g^2(\mu)}\right]. }$$

The same equation can be inverted:

$$g^2(\mu) = \frac{1}{2b_0\ln(\mu/\Lambda_{(1)})}, \qquad \mu>\Lambda_{(1)}.$$

For the exactly integrated two-term polynomial, an invariant is

$$\begin{aligned} \Lambda_{[2]} &= \mu \exp\left[-\frac{1}{2b_0g^2(\mu)}\right] \\ &\quad\times \left[ \frac{b_0g^2(\mu)}{ 1+(b_1/b_0)g^2(\mu) } \right]^{-b_1/(2b_0^2)}. \end{aligned}$$

The square brackets retain terms generated by integrating the truncated polynomial. In the small-coupling expansion, the conventional two-loop scale is

$$\boxed{ \Lambda_{(2)} = \mu \exp\left[-\frac{1}{2b_0g^2(\mu)}\right] \bigl[b_0g^2(\mu)\bigr]^{-b_1/(2b_0^2)} \left[1+O(g^2)\right]. }$$

Both forms are useful, but they answer slightly different questions. $\Lambda_{[2]}$ is exactly constant for the two-term polynomial. $\Lambda_{(2)}$ displays the universal small-$g$ structure and the size at which the omitted beta-function information enters. Collins gives the corresponding scale convention and asymptotic running in Collins 1984/2023, § 7.5.1, p. 188.

The characteristics figure packages this calculation geometrically. Inspect the lower panel: changing $\mu$ moves the representative point along one curve, while $\Lambda$ labels the curve and remains fixed.

RG characteristics convert the boundary datum $g(\mu_0)$ into a trajectory $g(\mu)$. In panel (c), $\beta=-b_0g^3$ with $b_0>0$ gives the invariant $\Lambda=\mu e^{-1/(2b_0g^2)}$. The diagram is schematic and not to scale; approaching $\mu\sim\Lambda$ signals loss of perturbative control, not a plotted physical singularity.

Figure element	Mathematical statement	Independent check
Characteristic through $(\mu_0,g_0)$	$dg/d\ln\mu=\beta(g)$ with one boundary value	Differentiate the integrated solution and recover $\beta(g)$
Constant label on the curve	$d\Lambda/d\ln\mu=0$	Substitute the running $g(\mu)$ into $\Lambda_{(1)}$
Infrared end of the weak-coupling segment	$\mu/\Lambda$ is no longer large	Stop before the chosen perturbativity criterion fails

What dimensional transmutation does—and does not—say

Consider a massless renormalized theory with one dimensionless coupling. The pair $(\mu_0,g_0)$ appears to contain a scale and a coupling, but changing $\mu_0$ while evolving $g_0$ along the same characteristic changes neither the bare theory nor a physical prediction. The invariant information can instead be written as the single dimensionful parameter $\Lambda$:

$$(\mu_0,g_0) \quad\longleftrightarrow\quad \Lambda$$

within a declared scheme, coupling normalization, perturbative order, and flow branch. At one loop,

$$\frac{1}{g_0^2} = 2b_0\ln\frac{\mu_0}{\Lambda_{(1)}}.$$

This is a one-to-one change of boundary coordinates, not an extra prediction. The exponential relation also creates a hierarchy:

$$\frac{\Lambda_{(1)}}{\mu_0} = \exp\left[-\frac{1}{2b_0g_0^2}\right] \ll1$$

when $g_0^2\ll1$. Correspondingly,

$$\delta\ln\Lambda_{(1)} = \frac{\delta g_0}{b_0g_0^3}$$

at fixed $\mu_0$ and $b_0$. Small changes in a weak reference coupling can therefore represent large multiplicative changes in the transmuted scale.

If a multiplicatively defined physical quantity $Q$ has mass dimension $d_Q$, no other dimensional input, and no additive local ambiguity, RG invariance and dimensional analysis imply

$$Q=C_Q\Lambda^{d_Q},$$

or, with an external momentum $p$,

$$Q(p)=p^{d_Q}\, \mathcal F_Q\left(\frac{p}{\Lambda}\right).$$

The dimensionless constant $C_Q$ and the function $\mathcal F_Q$ need not be perturbatively calculable near $p\sim\Lambda$. Perturbative running determines how the weak-coupling boundary datum is encoded in $\Lambda$; it does not generally determine a mass gap, bound-state spectrum, or confinement. When nonzero masses, relevant couplings, or thresholds are present, they supply additional invariant ratios and matching conditions.

This interpretation, including the fact that physical mass ratios are dimensionless constants while the common scale is transmuted, is developed in Collins 1984/2023, § 7.9, pp. 203–206.

Scheme and order labels

The numerical value of $\Lambda$ is not scheme independent. For a finite one-coupling redefinition

$$g'=g+a g^3+O(g^5),$$

the first two beta-function coefficients are unchanged within mass-independent schemes, but the one-loop-normalized scales obey

$$\Lambda' = e^{a/b_0}\Lambda.$$

That constant rescaling is not a physical contradiction. A matched dimensionless prediction, such as $M/\Lambda$ evaluated in the same scheme, transforms so that the observable is unchanged. The next page derives these finite-redefinition rules in detail.

A reproducible scale statement therefore has the form

Required label	Example of what must be stated
Coupling definition	normalization of $g$ and the renormalized vertex or observable defining it
Subtraction scheme	for example, a declared mass-independent minimal-subtraction convention
Beta-function order	one term, two terms, or a higher declared truncation
Active degrees of freedom	particle content used in $b_0,b_1,\ldots$ over the stated interval
Reference condition	$g(\mu_0)=g_0$ or an equivalent matched value of $\Lambda$
Validity and stopping rule	coupling bound, threshold, fixed point, or other event that terminates the evolution

Analytic benchmark for RG-flow computation

A reproducible calculation begins with a one-coupling test whose exact answer is known. Its default equation and boundary data are

$$\frac{dg}{dt}=-g^3, \qquad t=\ln\frac{\mu}{\mu_0}, \qquad g(0)=0.4.$$

Thus

$$g(t) = \frac{0.4}{\sqrt{1+0.32t}}, \qquad t>-3.125,$$

and the diagnostic invariant is

$$I(t) \equiv \frac{1}{g^2(t)}-2t = \frac{1}{g^2(0)} =6.25.$$

The exact reference values are:

$t$	$\mu/\mu_0$	$g(t)$	$I(t)$
$-2$	$0.135335283$	$0.666666666667$	$6.25$
$0$	$1$	$0.400000000000$	$6.25$
$2$	$7.38905610$	$0.312347523777$	$6.25$
$5$	$148.413159$	$0.248069469178$	$6.25$

A numerical solver should reproduce the invariant to relative tolerance $10^{-12}$, show $dg/dt<0$ on the positive branch, and reject evolution through $t=-3.125$. In a physical calculation the stronger stopping condition is normally reached earlier, when the coupling or the estimated omitted terms exceed the declared perturbative domain.

Regime limits and stopping events

An integrated RG trajectory is trustworthy only over the interval on which its inputs and approximation remain valid. Check the following events before interpreting the endpoint:

Event	Why the current solution must stop or change	Correct continuation
$g$ becomes too large	Omitted beta-function terms are no longer demonstrably smaller	Switch to controlled nonperturbative information or report loss of control
A mass threshold is crossed	The active fields and beta coefficients change	Match onto the appropriate theory and continue with its beta function
$\beta(g)$ reaches a reliable zero	The trajectory approaches a fixed point rather than a pole	Linearize about the zero and test stability and truncation dependence
A denominator in the truncated solution vanishes	The approximate ODE has left its weak-coupling branch	Do not continue through the formal singularity
A scheme map becomes singular or noninvertible	Coordinate equivalence has failed on that domain	Restrict to the invertible patch or choose a regular scheme
Large logarithms remain in the boundary calculation	Running alone did not choose all natural matching scales	Factorize or match at the relevant scales and evolve each object consistently

The threshold problem is developed in Decoupling Theorems and Threshold Corrections. Fixed-point stability belongs to Fixed Points and Linearized RG Flow, and the systematic treatment of large logarithms appears later in this chapter.

Common pitfalls

Calling $\Lambda$ scheme independent. It is RG invariant within a declared convention, but a finite coupling redefinition rescales it. Physical matched ratios are the invariant claims.

Equating a perturbative endpoint with a physical pole. A vanishing denominator says that the truncated weak-coupling solution has failed. It does not locate a pole of an exact observable.

Claiming that dimensional transmutation computes a mass gap. The RG converts $g(\mu_0)$ into $\Lambda$. A relation such as $M=C_M\Lambda$ still contains dimensionless dynamics that may be nonperturbative.

Forgetting the flow branch. The integral of $1/\beta(g)$ is branch dependent when beta functions have zeros or singularities. A boundary condition cannot be evolved through such a point by algebraic continuation alone.

Trusting an exact solution of a truncated beta function too far. Exact ODE integration removes numerical integration error; it does not remove perturbative truncation error.

Exercises

For $dg/dt=-b_0g^3$, show directly that $\Lambda_{(1)}=\mu e^{-1/(2b_0g^2)}$ is invariant and derive the scale at which a trajectory with boundary value $g_0$ leaves the one-loop branch.

Solution

Differentiate along the flow:

$$\frac{d\ln\Lambda_{(1)}}{dt} = 1+ \frac{1}{b_0g^3}\frac{dg}{dt} =1-1=0.$$

The integrated solution is $1/g^2=1/g_0^2+2b_0t$. Its denominator reaches zero at $t=-1/(2b_0g_0^2)$, hence $\mu=\mu_0e^{-1/(2b_0g_0^2)}=\Lambda_{(1)}$. Perturbation theory must be stopped before that formal endpoint.

Let $g'=g+a g^3+O(g^5)$. Using only the one-loop definition of $\Lambda$, derive the ratio $\Lambda'/\Lambda$.

Solution

The inverse coupling transforms as

$$\frac{1}{g'^2} = \frac{1}{g^2}-2a+O(g^2).$$

Therefore, in the $g\to0$ limit,

$$\begin{aligned} \ln\frac{\Lambda'}{\Lambda} &= -\frac{1}{2b_0g'^2} +\frac{1}{2b_0g^2} +O(g^2) \\ &= \frac{a}{b_0}, \end{aligned}$$

so $\Lambda'/\Lambda=e^{a/b_0}$. The finite factor must be included when translating any quantity quoted in units of $\Lambda$.

Where to continue

• Scheme Transformations and RG Invariants derives which beta-function data and fixed-point statements survive finite coordinate changes.
• Multiple Couplings and Coupled RG Flows replaces the one-dimensional trajectory by a vector field with stability matrices and separatrices.
• Large Logarithms and RG Improvement uses characteristic evolution to reorganize logarithmic perturbation theory.

References

• Collins, John C. Renormalization: An Introduction to Renormalization, the Renormalization Group, and the Operator-Product Expansion. Cambridge Monographs on Mathematical Physics. Cambridge: Cambridge University Press, 1984; open-access digital edition, 2023. DOI. Open PDF.